Wednesday, February 17, 2010

A Constraint Based Approach To Figure Skating

While perhaps not quite a pure crash blossom, this headline caught me off guard:

Honestly, my first reaction was to wonder if there was a new scoring system (yes, there is) and what was wrong with the old one (bias and collusion). In other words, what was broken and how was it improved? Of course, there's another meaning of fixed -- 'to cheat.'  In other words, are figure skating outcomes rigged by cheating?  Were this headline from any other publication than the increasingly dumbed down Slate, I'd assume the ambiguity was intentional, but with Slate these days, you just never know. Note that there are at least two other senses for the word fixed: to spay/neuter a pet and to have sufficient amount of something like money (British English as in 'You Kev mate, you fixed for goin' out later? HT Urban Dictionary). With at least 4 senses to choose from, no wonder I was a tad confused.

But how did my super duper human language processing system resolve this?

This headline reminded me of James Pustejovsky's somewhat older work on lexical ambiguity (aka polysemy) and the mechanism of lexical underspecification. For example, in The Semantics of Lexical Underspecification (1998) Pustejovsky contrasts the various senses of good:
  1. a good book
  2. a good meal
  3. a good knife
The adjective good does not mean the same thing in these three phrases because a book is good in a different sense than a knife is good.  The qualities that make a book good are mental while the qualities that make a knife good are physical. There is, however, some core meaning of good that's consistent across its different senses. In Pustejovsky's words, the adjective good "can be analyzed as an event modifier which subselects for a relational interpretation available in the head noun." In lay terms, he's proposing that each noun (book, meal, knife) carries with it as part of its meaning some notion of the kinds of things people normally do with them. So, people normally read books, eat meals, and cut with knives. If the meaning of the nouns (book, meal, knife) includes these events, then the adjective good could be modifying the events (the verbs read, eat, cut), not the noun. Note that this cuts against the grade school definition of adjectives as words that modify nouns. 

To work, this hypothesis requires a very complex lexical definition stored in the mental lexicon (the brain dictionary). Here's what the definition for book would have to look like:

For our purposes, we simply need to see that the lexical entry for good contains a TELIC feature read. When the noun book is modified by the adjective good, so says this hypothesis, the adjective is modifying the act of reading, not the book itself. Since each noun (book, meal, knife) has a different set of TELIC event features unique to them (read, eat, cut), the polysemy of good can be explained rather elegantly. The adjective good isn't necessarily polysemic at all; rather, the word good picks out something different depending on the noun it's modifying. It's the nouns that differ, not the adjective.

Back when I was a grad student, I really liked Pustejovsky's work, as well as HPSG, which bears a lot of similarity (Stanford has a nice page describing the Leading Ideas of HPSG, first amongst them is "Strict Lexicalism"). But I always had a sneaking suspicion that it was more engineering than science. In other words, there was a tremendous amount of literature explaining how a lexicalist system could account for a large variety of language phenomena, and precious little literature on whether or not there was any psychological reality to any of it. This suspicion was part of my own evolution into a very psycholinguistics oriented linguist (i.e., I mostly want to know how the brain does language). I'll note that this is a bit unfair for two reasons:
  1. There actually is some psycholinguistics research on HPSG and lexicalist approaches, just not as much as I'd like. Maybe my expectations are unfair but I spent a few minutes searching through Stanford's lengthy  HPSG Bibliography and found only two citations that looked like experimental tests of HPSG hypotheses (see here, and here). I didn't search thoroughly though.
  2. It's really really hard to design psycholinguistic experiments to test these hypotheses.Until the tools and field of neuroscience progress further, there's not much syntacticians can do about this.
Neither of these reasons are unique to HPSG either. Grammatical theory is hard to test in general because we just don't have the tools and understanding of the brain necessary to thoroughly investigate all the complex hypotheses. Nonetheless, as a student I tired of reading yet-another-HPSG-analysis-of construction X in language Y papers.

In any case, the grammar game was fixed by Chomsky's thugs for decades when they took even the most tame non-dominant grammarians to the CFG/Transformational/P&P/Minimalist vet to be fixed so I'm happy to see any and all lively debate supported. I'll leave it to Ivan and Tom to fix HPSG as needed to account for new psycholinguistic data. Besides, I'm fixed with blogging topics for now, I don't need any more.

    James Pustejovsky (1998). The Semantics of Lexical Underspecification Folia Linguistica

    No comments:

    Putting the Linguistics into Kaggle Competitions

    In the spirit of Dr. Emily Bender’s NAACL blog post Putting the Linguistics in Computational Linguistics , I want to apply some of her thou...